AI SEARCH · 2026-10-30
AI Search — Paper Daily, 2026-10-30
Today's AI search track is about better retrieval not automatically giving better answers. One study decides per question whether to add a hybrid retrieval pass, keeping the evidence hit rate while skipping most extra calls, yet answers do not get more reliable. SpectRAG uses a chunk graph to fill in multi-hop evidence, with large recall gains but small answer gains. Two more papers tackle cost and risk: a government service assistant routes most requests to fixed templates and checks every amount, and MetaKV reuses recurring structures in GraphRAG to save prefill.
4.10
Is adding a second retrieval pass to RAG worth it? This paper first gets evidence with vector retrieval plus reranking, then has a small linear model read the scores left by that pass, such as how high the top-1 similarity is and how far it sits above the second result. From those it predicts whether switching to hybrid BM25-plus-vector retrieval would find more evidence, and it switches only when the predicted gain is large enough. On papers from QASPER (a QA dataset over research papers) that were never used in training, it switched only 24.5% of questions to hybrid retrieval and matched the evidence hit rate of always using hybrid retrieval, 78.70% for both. Yet when a small model generated answers from that evidence, neither correctness nor faithfulness improved. Finding evidence more cheaply does not make answers more reliable, and gains in retrieval metrics cannot be read as gains in answer quality.
3.80
For a government service-desk assistant, one invented fee is an incident, and citizens' questions are often vague or outside the agency's jurisdiction. The official portal of the Mexican state of Tlaxcala runs a RAG assistant on local servers. For each question it first decides whether to answer, to ask for the missing detail, or to refuse because the matter is outside the state's jurisdiction. After generating an answer, a program checks every fee and validity period against the retrieved official record. Across 380 real conversations after launch, 89.47% were answered with fixed templates, with no free generation at all; no invented amounts appeared in the audited generated answers. Where zero errors are tolerated, routing most requests away from free generation and adding a deterministic check to the rest is more robust than simply improving generation quality.
3.80
In multi-hop QA, retrieval recall jumps a lot while answers improve only a little. Offline, SpectRAG links document chunks into a graph: two chunks get an edge if they are semantically close, adjacent in the source, mention the same entities or cite each other, and these signals are weighted into the edge weight. At query time, vector retrieval and BM25 first hit some chunks, which then pass their relevance scores to neighbors along the edges, more to closer neighbors, so chunks that look unlike the question but are intermediate reasoning steps can be recovered. Finally it picks a low-redundancy context within about 12k tokens. On MultiHopRAG, its recall over the top 5 results is 64.7 versus 49.2 for the graph-retrieval baseline LightRAG. Yet its answer EM is only 1.5 points higher than LightRAG's. A large recall lead shrinks to a small edge at generation time, so multi-hop RAG should be judged on both retrieval and answers.
3.05
Every time GraphRAG answers a question, it flattens the retrieved subgraph into a long text and makes the LLM redo prefill. Yet while different questions retrieve different entities, the relational structure among them recurs again and again, so most compute goes into re-encoding the same structure. MetaKV mines these frequent structures in advance, turns each into a skeleton with empty slots, and precomputes the skeleton's KV cache. At inference it fetches the matching skeleton's cache, computes fresh representations only for this question's entities, and places them in the skeleton's reserved slots to assemble the full context. On HotpotQA and MetaQA, prefill is up to 6.4 times faster with a 73% effective cache-hit rate, while reasoning accuracy stays competitive. Entities change but the structure does not, so the structure should not be recomputed every time.
KDD 2026